<?xml version="1.0" standalone="yes"?>
<Paper uid="C94-2137">
  <Title>Fr6d~rique Segond</Title>
  <Section position="5" start_page="856" end_page="857" type="concl">
    <SectionTitle>
5. Conclusion
</SectionTitle>
    <Paragraph position="0"> We have compared two &amp;quot;alternative methods of compulalion of long-distance dependencies: the CCG and AUG methods. Both methods ,are consistent with respect to their mathematical machinery.</Paragraph>
    <Paragraph position="1"> The essential difference between the two methods is that while AUG with its theory of superposition expands its formalism to reflect file linguistic reality, CCG, by abandoning the normal constituency analysis, gets caught up in its formalism to lapse into linguistic unreality.</Paragraph>
    <Paragraph position="2"> CCG analysis produces phantoms, as: (He must) leave.</Paragraph>
    <Paragraph position="3"> ((He must) love) her. (31) This startling analysis does not permit us to correctly describe agreement, government and clitization. These artificial constituent structures are completely divorced from the syntactic and semantic reality.</Paragraph>
    <Paragraph position="4"> The CCG's use of type raising iu conjunction with type composition changes the initial natural types assigned to words into artificial types and produces artificial constituents for the convenience of computation. By contrast, supeq)osition, in conj unction with the Principle of Elhnination of Empty Constituents and the Principle of Syntactic Assimilation, changes natund types into natural types and produces syutactically and semantically appropriate constituents without any sacrifice in the consistency of the mathematical formalism or in the convenience of computation.</Paragraph>
    <Paragraph position="5"> In support of their departure from the accepted ealalyses of syntactic constituents the proponents of the CCG refer to psychological studies ou speech recognition claiming that hmnan &amp;quot;recognizer&amp;quot; works &amp;quot;from left to right&amp;quot;.(Ades and Stcedman, 1982: 517-518).</Paragraph>
    <Paragraph position="6"> ~\[kvo problems arise here. First, although human speech is linear and the words of a sentence are produced from left to right, so to say, that does not mean that the listener analyzes spccch word by word. It is reasonable to assume that the listener performs the analysis of a sentence first by syntactic blocks and then globally. There is no conchlsive psychological evidence that tile hearer's recognition of tile sentence sUucture corresponds to the CCG method that disposes with the normal constituency analysis.</Paragraph>
    <Paragraph position="7"> Second, psychological phenomena are irrelevant to confirmation or refutation of linguistic theories, because fin- null guistics is completely independent of psychology. True, linguistic processes involve the psychological processes in the human mind. But logical and mathematical reasoning also involve psychological processes, llowever, nobody tries to b~se logic or mathematics on psychology.</Paragraph>
    <Paragraph position="8"> Linguistics is part of semiotics--the theory of sign systems. Sign systems, as well as mathematical systems, are in the human mind. But the laws of semiotics and mathematics are different from the laws of psychology.</Paragraph>
    <Paragraph position="9"> One may argue that computational linguistics is different from ordinary linguistics ,and therefore any parser will do for computational linguistics as long as it &amp;quot;works&amp;quot;. We believe that good computatiolml linguistics must be good linguistics ,'tq well. Both ordinary and computational linguistics must share common theoretical principles characterizing the nature of human language. Computatioual linguistics is not second-rate linguistics where anything goes.The real difference between the two types of linguistics is that compuUltional linguistics exp,'mds ordinary linguistics by rules characterizing its interaction with computers rather than distorts it. Computational linguistics is at the cutting edge of the study of hum,'m lauguage: it must enrich our understauding of all its aspects, rather them fudge the linguistic concepts for the sake of the ease of the implementation.</Paragraph>
    <Paragraph position="10"> The irreparable defect of the CCG method is that it produces phantom constituents m~d phautom slructures that prechtde a correct analysis of linguistic processes.</Paragraph>
    <Paragraph position="11"> The CCG method is interesting attd important as an experiment in rite application of combiuators in linguistics. The negative results of this experiment ~u'e important in that they reveal the hazards involved in the use of combinators (for use of combiuators in AUG, see Shaumyan, 1987; Descl6s, 1990; Descl6s et al. 1985, 1986).</Paragraph>
    <Paragraph position="12"> As an instrument of cognition mathematics has a specific function--to be a tool of deduction. But deduction is neutral to file value of ideas. It is like a mill: if you put grain into it, you will get flour; ~utd if you put in chaff, you will get processed chaff. Mathematical consistency does uot guarantee a correct description of reality. &amp;quot;Side by side with mathematization of knowledge, mathematization of nonsense also goes on (N~dimov, 1981: 149).&amp;quot; The use of mathematics as a tool of deduction makes sense only when the initial ideas from which we deduce their consequences have value (on use and abuse of mathematical formalism, see Shaumy,'m 1987: 28-29, 318-321).</Paragraph>
    <Paragraph position="13"> In conclusion, we would like to say a few words about Ihe computer implementation of AUG, Fr6d6rique Segond has implemented AUG and its theory of superposition to deal with infinitive clauses and gerunds in French (for a complete description of the parser, see Segond, 1990a).</Paragraph>
    <Paragraph position="14"> This parser has been implemented in PLNLP (Program~ ruing Language for Natural Language Processing, described in lleidom, 1972) at the IBM Research Center in l'aris. The parser uses a machine dictionary of 50,000 ena'ies ~md was tested on more thm~ one hundred different types of sentences, including constructions such as relative clauses, simple cases of coordinatiou, infinitive clauses, and gerunds, among others. Currently Sebasli~m Shaumyan is working on implementing AUG in functional programming languages (Miranda, I Iaskell).</Paragraph>
  </Section>
class="xml-element"></Paper>